科研 Skill
AI对标专家基准研究设计
design-ai-benchmarking
Plans the arms, rubrics, calibration probes and reviewer panel for an AI-vs-expert benchmark. Use before data collection when designing an AI-vs-human rating study.
SKILL.md
Design-AI-Benchmarking Skill
Standard Output
## AI-Benchmark Design Review
Evaluation question: ...
Arms / systems compared: ...
Reference (human-expert panel): ...
Unit of rating: (item / case / output)
### Rubric (decoupled dimensions)
- dimension -> construct -> anchors (1..k)
### Calibration probes (blinded, randomized)
- positive-control / known-bad / instability / mechanism-contradiction
### Reviewer panel
- n reviewers, metadata captured, per-reviewer randomized order
### Reliability plan
- IRR target on the anchor set of real items (ICC form + 95% CI) + control-item hit rate (a competence check, reported separately)
### Judge strategy
- human-as-judge / LLM-as-judge / both + adjudication rule
### Validity risks
1. ...
### Minimal fixes
- ...
### Decision
- Ready to collect / Needs rubric revision / Needs arm or judge redesign
Reviewer ratings, reference labels, probe outcomes and agreement statistics come only from collected
rating records. Never invent them: a reported ICC, kappa or score with no underlying rating record is
the failure this skill exists to prevent. Cite a reference only with a /search-lit-confirmed DOI or
PMID; mark any other [UNVERIFIED - NEEDS MANUAL CHECK]. Flag an unconfirmed clinical definition,
diagnostic criterion or guideline recommendation [VERIFY] and ask the user.
Workflow
Phase 1: Define the evaluation question and arms
Pin down, in writing:
- the exact claim the benchmark must support (e.g., "system A's outputs are perceptually indistinguishable from expert outputs", not "system A is deployment-ready")
- every arm/system and what each receives as input (same items, same information access, same output format), so no arm has a hidden advantage
- the human-expert reference: who they are, and whether they set ground truth, form a comparison arm, or both
- the unit of rating (item, case, output) and how many units each reviewer sees
Gate: Present the reconstructed evaluation question, arms, and reference to the user and confirm before designing the rubric. A wrong reconstruction misdirects the entire benchmark.
Phase 2: Design a decoupled multi-dimensional rubric
- Decouple the axes. Each rated dimension measures one construct. Keep "is the output valid/correct" separate from "is it novel", "is it feasible/measurable", "does it add value over current tools", and "would it change action". A candidate can be high-validity yet low-added-value ("real but redundant"); a single blended score hides this.
- Anchor every scale point with a short verbal descriptor; pilot the anchors with at least one reviewer before locking.
- Pre-specify discriminant validity: hypothesize which dimensions should correlate vs be orthogonal, then report the full inter-dimension correlation matrix.
- Start from
references/elicitation_rubric_template.md(dimensions, anchors, probe flavors).
Phase 3: Insert and randomize calibration probes
Plant a few deliberate control items, blinded and randomized across raters (record who received which
via a probe_arm flag), to anchor the scale, measure rater drift/fatigue, and audit the rubric and
pipeline. Four flavors:
- Positive control / "too-good" item — near-tautological; tests whether raters equate "largest effect" with "best", and whether the construct-independence gate (Phase 7) works.
- Known-bad negative control — an engineered defect (fabricated reference, missing key statistic).
- Instability item — an estimate that reverses or fails to replicate on a holdout.
- Mechanism-contradiction item — an empirical direction that opposes the proposed mechanism.
Probes are planted or adjudicated, never fabricated to fit a hypothesis.
Phase 4: Construct the reviewer panel
- Recruit reviewers spanning the intended expertise gradient; pre-specify any stratification.
- Capture reviewer metadata (years of experience, prior AI-evaluation experience, subspecialty).
- Randomize item order per reviewer (not one global seed), record the order, and plan to analyze order and fatigue effects.
- Require each item to be judged standalone; cross-item references in free text signal non-independent rating.
Gate: Present the panel composition, stratification, and randomization plan for user review before recruitment is finalized.
Phase 5: Set inter-rater reliability targets
- Pre-specify the reliability statistic and its form — for several raters, an ICC with the model stated (e.g. ICC(2,1): two-way random, absolute agreement, single rater) or Krippendorff's α (ordinal); weighted (Cohen's) kappa is a two-rater statistic — and a justified target. Size the number of items for the CI width you need: ICC precision is driven mainly by the item count (Koo & Li, J Chiropr Med 2016).
- Planted control items are a rater-competence check, not reliability evidence. Report the hit rate on known-good vs known-bad controls. Do not present an ICC computed on the controls as evidence of rubric or scale validity: ICC scales with the spread between items, controls are extreme by construction, so their ICC is high whatever the raters' reliability on real items (same five raters, same noise: ICC(2,1) 0.89 on four planted controls, 0.16 on 40 real items), and it rests on a handful of items.
- Report reliability on the representative anchor set of real items — ICC with its 95% CI and stated form — with an agreement measure alongside (exact agreement, SEM), since reliability depends on how heterogeneous the rated items are (de Vet et al., J Clin Epidemiol 2006). A low ICC on real items with a good control hit rate points to little true spread among the items or a rubric that does not separate them, not to careless raters.
- Plan the minimum ratings-per-item for a stable agreement estimate (the math goes to
/analyze-stats).
Phase 5b: Reader allocation under burden constraints (anchor-and-rotate)
When the item pool exceeds what one reader can rate in a session, do not make every reader rate
every item (that caps the pool at the per-reader limit). Use anchor-and-rotate (an anchor set
plus rotating incomplete blocks — not a balanced incomplete block design, since reader pairs are
not co-rated equally often): all readers rate a shared anchor set of real items (which carries
the inter-rater reliability; the planted controls carry only the competence check) plus a
rotating unique block each. The binding
constraint is usually the number of available expert readers, so solve the reverse problem (largest
pool for R readers) to size the must-rate set. Pre-specify anchor membership, raters-per-item, and the
rotation seed before rating. Read references/anchor_rotate_reader_allocation.md
for the formulas, trade-offs, and a stdlib implementation.
Phase 6: Choose the judge strategy and adjudication
- Decide human-as-judge, LLM-as-judge, or both. An LLM judge is one more arm whose ratings must be validated against the human panel on the anchor set of real items, not only on the extreme control items.
- Pre-specify the adjudication rule for disagreement (majority, a third senior reviewer, consensus discussion) and who adjudicates.
- Blind judges to arm identity wherever feasible; record any unavoidable unblinding.
Phase 7: Construct-independence and leakage guards
- Exclude any predictor or input that is a definitional component of the outcome, and flag near-tautological composites built from the outcome's defining components — they produce an inflated, near-circular result and belong as labeled probes, not discoveries.
- Verify no arm sees post-decision or outcome-derived information the others do not.
- Confirm the reference labels were not derived from the same model output being evaluated.
Phase 8: Lock a structured export schema
Write the machine-readable rating record as a JSON schema, starting from
references/benchmark_export_schema.json: per-item ratings on every rubric
dimension, free-text justifications, follow-up flags, the probe_arm flag, reviewer id and metadata,
item order, and timing.
Gate: Present the final rubric, probe set, panel plan, judge strategy, and export schema together; collect explicit user approval before any rating begins, because changes after collection starts compromise the comparison.
Handoff Rules
- route to
/analyze-statsfor ICC (stated form) / Krippendorff's α / kappa (two raters) / DeLong, agreement sample size, and effect-size real-world translation of the benchmark results - route to
/check-reportingfor STARD-AI, CLAIM, or TRIPOD+AI item-level reporting once the design is locked - route to
/design-studywhen the broader study around the benchmark (cohort logic, analysis unit, comparator) also needs review - route to
/peer-reviewor/self-reviewonly after ratings exist and a manuscript is being assessed
示例提示词
帮我设计一个AI系统对比专家医生的基准测试研究方案。
所属插件
安装该插件会一并获得同领域的 58 个 Skill / MCP。
同插件的其他科研 Skill
- 医学AI论文检索优化Optimizes a medical AI paper’s title, abstract and metadata for AI search engines. Use when a paper should be findable and citable by RAG tools and AI search.
- 临床数据统计分析与建模Generates reproducible Python or R code for clinical statistics and tables. Use when collected data needs Table 1, regression, survival or agreement analysis.
- 作者发表记录与策略分析Analyzes a PubMed author record to classify study types and authorship patterns. Use when planning a publication strategy from a researcher’s track record.
- 批量队列暴露结局分析Generates one cohort-analysis script per exposure-outcome pair from one template. Use when a validated method must run across many variable combinations at once.
- 样本量估算与检验效能Walks a decision tree to the right test and computes sample size with IRB text. Use when planning how many patients a study needs before data collection.
- 因果推断方法设计Runs DAG-based, instrumental-variable, difference-in-differences or RD analyses. Use when an association in observational data must support a causal claim.
- 报告规范逐条核查Audits a manuscript item by item against 49 reporting and risk-of-bias checklists. Use when checking a draft against STROBE, CONSORT, TRIPOD or similar guidelines.
- 临床数据清洗与画像Profiles a clinical dataset and generates cleaning code in staged, approved steps. Use when a CSV or Excel dataset needs missing-value and outlier checks before analysis.
- 跨国队列数据协调对比Harmonizes variables and runs parallel weighted analyses across national surveys like KNHANES and NHANES. Use when comparing one exposure-outcome link across countries.
- 变量定义与截断值溯源Maps exposure, outcome and covariate definitions to a cited guideline and DB columns. Use when variable cutoffs need a citable basis before the study protocol.
- 本地化临床数据脱敏Runs a local, network-free script that detects and removes identifiers from clinical data. Use when data may contain PHI and must be de-identified before LLM analysis.
- 放射影像AI研究设计审查Reviews a study’s analysis unit, cohort logic, leakage risk and reporting-guideline fit. Use when checking a clinical or medical-AI study design before drafting.
- 传染病仓室模型拟合Builds and fits SIR, SEIR and SEIHR compartmental models with Rt estimation. Use when an outbreak or epidemic’s spread needs mathematical modeling in Python.
- ICMJE利益冲突声明填写Clones a pre-filled synthetic ICMJE COI form, replacing only date, name and title. Use when each author on a submission needs a disclosure form filled quickly.
- 机构Word表单内容填充Fills an institutional Word form, such as an IRB protocol, without breaking its layout. Use when drafted content must be rendered into a CJK-aware .docx template.
- 队列数据库选题挖掘Profiles a longitudinal cohort database and ranks research-topic proposals with gap evidence. Use when looking for a question an available cohort or registry can answer.
- 投稿期刊匹配推荐Matches a manuscript abstract against curated journal profiles and ranks candidates. Use when choosing where to submit, or re-targeting after a rejection.
- DOI全文批量获取Batch-downloads open-access PDFs for a list of DOIs via Unpaywall, PMC and OpenAlex. Use when full texts are needed for a screening set such as a meta-analysis.
- 数据集编码本自动生成Profiles every variable in a tabular dataset into a structured, citable codebook. Use when a CSV, Excel or Stata dataset needs a data dictionary before analysis.
- 全球卫生数据获取分析Accesses and analyzes global health data from WHO GHO, IHME GBD and Our World in Data. Use when DALYs, age-standardized rates or choropleth maps are needed.
- 医学AI项目基金申请书撰写Structures a grant proposal’s significance, innovation, approach, aims and milestones. Use when drafting a research funding proposal, including a Korean grant plan.
- 卫生经济学评价建模Builds Markov cohort models with QALYs, PSA and cost-effectiveness planes. Use when a health intervention needs a formal cost-effectiveness evaluation.
- AI写作痕迹识别改写Scans a manuscript for AI writing patterns and rewrites flagged passages carefully. Use when a draft or response letter reads as obviously AI-written.
- 医学影像数据集画像与预处理Profiles a DICOM or NIfTI dataset’s spacing, intensity and label integrity before modeling. Use when preparing an imaging dataset and auditing it for leakage.
常见问题
AI对标专家基准研究设计能做什么?
Plans the arms, rubrics, calibration probes and reviewer panel for an AI-vs-expert benchmark. Use before data collection when designing an AI-vs-human rating study。
如何安装AI对标专家基准研究设计?
在科搭 Codar 左侧点击「插件」,在「组件」页签搜索「AI对标专家基准研究设计」并安装;在 Claude Code 或 Codex 中可以发送本页的安装提示词,或下载 Skill 包后解压到 ~/.claude/skills 或 ~/.codex/skills。
AI对标专家基准研究设计需要哪些权限?
该科研 Skill没有声明额外权限;执行具体操作时仍由 Agent 按自身规则请求确认。
AI对标专家基准研究设计经过官方审核吗?
是。SpeedSkills 收录的每个版本都会记录内容摘要与权限声明,当前版本已通过安全审核。